Why compute the likelihood contributed by different scales in this way?

[williacode]

I have two questions about the likelihood part.
Question 1:
In graph.py,

hierarchical_anomaly_detection/invglow/invertible/graph.py

Line 222 in ca2f1d8

return xs[self.index], sum(prev_log_dets) / n_parts

the logdet is divided by n_parts (usually 2).
Why the logdet is divided by 2 at SelectNode?
Question 2:
Why the likelihood contributed by different scales is computed in the following way?
Here

hierarchical_anomaly_detection/invglow/evaluate.py

Line 38 in ca2f1d8

lp0 = node_to_lps['m0-flow-0'] / 2 + node_to_lps['m0-dist-0']

lp0 is computed as the likelihood contributed by the first scale.
and here

hierarchical_anomaly_detection/invglow/evaluate.py

Line 41 in ca2f1d8

lp2 = node_to_lps['m0-dist-2'] - node_to_lps['m0-flow-1'] / 2

lp2 is computed as the likelihood contributed by the last scale.
But I have checked the formula (3) in the paper 'Understanding Anomaly Detection with Deep Invertible Networks through Hierarchies of Distributions and Features'.
The likelihoods contributed by different scales are not equal to the implementation.
Can somebody explain this?